Implement simple OpenMP breakdown view

cfg/cpu/openmp/breakdown.cfg

@@ -0,0 +1,44 @@
+#ParaverCFG
+ConfigFile.Version: 3.4
+ConfigFile.NumWindows: 1
+
+
+################################################################################
+< NEW DISPLAYING WINDOW CPU: OpenMP Runtime/Label breakdown >
+################################################################################
+window_name CPU: OpenMP Runtime/Label breakdown
+window_type single
+window_id 1
+window_position_x 100
+window_position_y 100
+window_width 600
+window_height 150
+window_comm_lines_enabled false
+window_flags_enabled false
+window_noncolor_mode true
+window_custom_color_enabled true
+window_custom_color_palette {1.000000000000:0,70,0},{3.000000000000:99,131,0},{4.000000000000:53,121,221},{5.000000000000:223,108,0},{6.000000000000:75,127,82},{7.000000000000:242,110,162},{8.000000000000:255,190,0},{9.000000000000:153,114,0},{10.000000000000:156,12,231},{11.000000000000:177,25,229},{12.000000000000:255,72,50},{15.000000000000:0,171,255},{16.000000000000:124,213,228},{17.000000000000:242,239,141},{18.000000000000:80,80,80},{19.000000000000:94,0,0},{20.000000000000:128,165,214},{22.000000000000:222,101,128},{23.000000000000:110,148,255},{100.000000000000:0,100,0},{101.000000000000:100,100,177},{102.000000000000:150,150,0}
+window_logical_filtered true
+window_physical_filtered false
+window_comm_fromto true
+window_comm_tagsize true
+window_comm_typeval true
+window_units Microseconds
+window_maximum_y 1000.0
+window_minimum_y 1.0
+window_compute_y_max true
+window_level thread
+window_scale_relative 1.000000000000
+window_end_time_relative 1.000000000000
+window_object appl { 1, { All } }
+window_begin_time_relative 0.000000000000
+window_open true
+window_drawmode draw_randnotzero
+window_drawmode_rows draw_randnotzero
+window_pixel_size 1
+window_labels_to_draw 1
+window_selected_functions { 14, { {cpu, Active Thd}, {appl, Adding}, {task, Adding}, {thread, Last Evt Val}, {node, Adding}, {system, Adding}, {workload, Adding}, {from_obj, All}, {to_obj, All}, {tag_msg, All}, {size_msg, All}, {bw_msg, All}, {evt_type, =}, {evt_value, All} } }
+window_compose_functions { 9, { {compose_cpu, As Is}, {compose_appl, As Is}, {compose_task, As Is}, {compose_thread, As Is}, {compose_node, As Is}, {compose_system, As Is}, {compose_workload, As Is}, {topcompose1, As Is}, {topcompose2, As Is} } }
+window_filter_module evt_type 1 53
+window_filter_module evt_type_label 1 "CPU: OpenMP Runtime/Label breakdown"
+

src/emu/CMakeLists.txt

@@ -65,6 +65,7 @@ add_library(emu STATIC
   kernel/event.c
   openmp/setup.c
   openmp/event.c
+  openmp/breakdown.c
 )
 target_link_libraries(emu ovni-static)
 

src/emu/emu_prv.h

@@ -35,6 +35,7 @@ enum emu_prv_types {
 	PRV_OPENMP_SUBSYSTEM = 50,
 	PRV_OPENMP_LABEL     = 51,
 	PRV_OPENMP_TASKID    = 52,
+	PRV_OPENMP_BREAKDOWN = 53,
 	PRV_OVNI_MARK        = 100,
 	/* User marks [100, 200) */
 	PRV_RESERVED         = 200,

src/emu/nanos6/setup.c

@@ -313,7 +313,7 @@ model_nanos6_create(struct emu *emu)
 	extend_set(&emu->ext, model_id, e);
 
 	if (model_nanos6_breakdown_create(emu) != 0) {
-		err("model_nanos6_breakdown_connect failed");
+		err("model_nanos6_breakdown_create failed");
 		return -1;
 	}
 

src/emu/openmp/breakdown.c

@@ -0,0 +1,239 @@
+/* Copyright (c) 2024 Barcelona Supercomputing Center (BSC)
+ * SPDX-License-Identifier: GPL-3.0-or-later */
+
+#include "breakdown.h"
+#include <stdint.h>
+#include <stdio.h>
+#include "bay.h"
+#include "chan.h"
+#include "common.h"
+#include "cpu.h"
+#include "emu.h"
+#include "emu_args.h"
+#include "emu_prv.h"
+#include "extend.h"
+#include "model_cpu.h"
+#include "mux.h"
+#include "openmp_priv.h"
+#include "proc.h"
+#include "pv/pcf.h"
+#include "pv/prf.h"
+#include "pv/prv.h"
+#include "pv/pvt.h"
+#include "recorder.h"
+#include "sort.h"
+#include "system.h"
+#include "task.h"
+#include "track.h"
+#include "value.h"
+
+
+static int
+create_cpu(struct bay *bay, struct breakdown_cpu *bcpu, int64_t gindex)
+{
+	enum chan_type t = CHAN_SINGLE;
+	chan_init(&bcpu->tri, t, "openmp.cpu%"PRIi64".breakdown.tri", gindex);
+
+	if (bay_register(bay, &bcpu->tri) != 0) {
+		err("bay_register tri failed");
+		return -1;
+	}
+
+	return 0;
+}
+
+int
+model_openmp_breakdown_create(struct emu *emu)
+{
+	if (emu->args.breakdown == 0)
+		return 0;
+
+	struct openmp_emu *memu = EXT(emu, 'P');
+	struct breakdown_emu *bemu = &memu->breakdown;
+
+	/* Count phy cpus */
+	struct system *sys = &emu->system;
+	int64_t nphycpus = (int64_t) (sys->ncpus - sys->nlooms);
+	bemu->nphycpus = nphycpus;
+
+	/* Create a new Paraver trace */
+	struct recorder *rec = &emu->recorder;
+	bemu->pvt = recorder_add_pvt(rec, "openmp-breakdown", (long) nphycpus);
+	if (bemu->pvt == NULL) {
+		err("recorder_add_pvt failed");
+		return -1;
+	}
+
+	if (sort_init(&bemu->sort, &emu->bay, nphycpus, "openmp.breakdown.sort") != 0) {
+		err("sort_init failed");
+		return -1;
+	}
+
+	for (struct cpu *cpu = sys->cpus; cpu; cpu = cpu->next) {
+		if (cpu->is_virtual)
+			continue;
+
+		struct openmp_cpu *mcpu = EXT(cpu, 'P');
+		struct breakdown_cpu *bcpu = &mcpu->breakdown;
+
+		if (create_cpu(&emu->bay, bcpu, cpu->gindex) != 0) {
+			err("create_cpu failed");
+			return -1;
+		}
+	}
+
+	return 0;
+}
+
+static int
+select_label(struct mux *mux, struct value value, struct mux_input **input)
+{
+	dbg("selecting tri output for value %s", value_str(value));
+
+	if (value.type != VALUE_NULL) {
+		dbg("selecting input 0 (label)");
+		*input = mux_get_input(mux, 0);
+	} else {
+		dbg("selecting input 1 (subsystem)");
+		*input = mux_get_input(mux, 1);
+	}
+
+	return 0;
+}
+
+static int
+connect_cpu(struct bay *bay, struct openmp_cpu *mcpu)
+{
+	struct breakdown_cpu *bcpu = &mcpu->breakdown;
+
+	/* Channel aliases */
+	struct chan *subsystem = &mcpu->m.track[CH_SUBSYSTEM].ch;
+	struct chan *label = &mcpu->m.track[CH_LABEL].ch;
+	struct chan *tri = &bcpu->tri;
+
+	/* Connect mux using label as select */
+	if (mux_init(&bcpu->mux, bay, label, tri, select_label, 2) != 0) {
+		err("mux_init failed");
+		return -1;
+	}
+
+	if (mux_set_input(&bcpu->mux, 0, label) != 0) {
+		err("mux_set_input label failed");
+		return -1;
+	}
+
+	if (mux_set_input(&bcpu->mux, 1, subsystem) != 0) {
+		err("mux_set_input subsystem failed");
+		return -1;
+	}
+
+	/* Emit unknown subsystem on NULL */
+	//mux_set_default(&bcpu->mux, value_int64(ST_UNKNOWN_SS));
+
+	return 0;
+}
+
+int
+model_openmp_breakdown_connect(struct emu *emu)
+{
+	if (emu->args.breakdown == 0)
+		return 0;
+
+	struct openmp_emu *memu = EXT(emu, 'P');
+	struct breakdown_emu *bemu = &memu->breakdown;
+	struct bay *bay = &emu->bay;
+	struct system *sys = &emu->system;
+
+	int64_t i = 0;
+	for (struct cpu *cpu = sys->cpus; cpu; cpu = cpu->next) {
+		if (cpu->is_virtual)
+			continue;
+
+		struct openmp_cpu *mcpu = EXT(cpu, 'P');
+		struct breakdown_cpu *bcpu = &mcpu->breakdown;
+
+		/* Connect tri channels and muxes */
+		if (connect_cpu(bay, mcpu) != 0) {
+			err("connect_cpu failed");
+			return -1;
+		}
+
+		/* Connect tri to sort */
+		if (sort_set_input(&bemu->sort, i, &bcpu->tri) != 0) {
+			err("sort_set_input failed");
+			return -1;
+		}
+
+		/* Connect out to PRV */
+		struct prv *prv = pvt_get_prv(bemu->pvt);
+		long type = PRV_OPENMP_BREAKDOWN;
+		long flags = PRV_SKIPDUP;
+
+		/* We may emit zero at the start, when an input changes and all
+		 * the other sort output channels write a zero in the output,
+		 * before the last value is set in prv.c. */
+		flags |= PRV_ZERO;
+
+		struct chan *out = sort_get_output(&bemu->sort, i);
+		if (prv_register(prv, (long) i, type, bay, out, flags)) {
+			err("prv_register failed");
+			return -1;
+		}
+
+		i++;
+	}
+
+	return 0;
+}
+
+int
+model_openmp_breakdown_finish(struct emu *emu,
+		const struct pcf_value_label **labels)
+{
+	if (emu->args.breakdown == 0)
+		return 0;
+
+	struct openmp_emu *memu = EXT(emu, 'P');
+	struct breakdown_emu *bemu = &memu->breakdown;
+	struct pcf *pcf = pvt_get_pcf(bemu->pvt);
+	long typeid = PRV_OPENMP_BREAKDOWN;
+	char label[] = "CPU: OpenMP Runtime/Label breakdown";
+	struct pcf_type *pcftype = pcf_add_type(pcf, (int) typeid, label);
+	const struct pcf_value_label *v = NULL;
+
+	/* Emit subsystem values */
+	for (v = labels[CH_SUBSYSTEM]; v->label; v++) {
+		if (pcf_add_value(pcftype, v->value, v->label) == NULL) {
+			err("pcf_add_value ss failed");
+			return -1;
+		}
+	}
+
+	/* Emit label values */
+	struct system *sys = &emu->system;
+	for (struct proc *p = sys->procs; p; p = p->gnext) {
+		struct openmp_proc *proc = EXT(p, 'P');
+		struct task_info *info = &proc->task_info;
+		if (task_create_pcf_types(pcftype, info->types) != 0) {
+			err("task_create_pcf_types failed");
+			return -1;
+		}
+	}
+
+	/* Also populate the row labels */
+	struct prf *prf = pvt_get_prf(bemu->pvt);
+	for (int64_t row = 0; row < bemu->nphycpus; row++) {
+		char name[128];
+		if (snprintf(name, 128, "~CPU %4" PRIi64, bemu->nphycpus - row) >= 128) {
+			err("label too long");
+			return -1;
+		}
+
+		if (prf_add(prf, (long) row, name) != 0) {
+			err("prf_add failed for %s", name);
+			return -1;
+		}
+	}
+
+	return 0;
+}

src/emu/openmp/breakdown.h

@@ -0,0 +1,48 @@
+/* Copyright (c) 2024 Barcelona Supercomputing Center (BSC)
+ * SPDX-License-Identifier: GPL-3.0-or-later */
+
+#ifndef BREAKDOWN_H
+#define BREAKDOWN_H
+
+/* 
+ * The breakdown model is implemented on top of the CPU subsystem, task_type and
+ * idle channels. The first mux0 selects the task type when the subsystem
+ * matches "Task body" otherwise forwards the subsystem as-is to tr. The second
+ * mux1 selects tr only when the CPU is not Idle, otherwise sets the output tri
+ * as Idle.
+ *
+ *            +--------+
+ *            |        |
+ *            |        v
+ *            |     +------+
+ * label -----+-->--|      |
+ *                  | mux0 |
+ * subsystem ---->--|      |-->-- tri
+ *                  +------+
+ * Then the sort module takes the output tri of each CPU and sorts the values
+ * which are propagated to the PRV directly.
+ *
+ *                 +------+       +-----+
+ * cpu0.tri --->---|      |--->---|     |
+ * ...             | sort |  ...  | PRV |
+ * cpuN.tri --->---|      |--->---|     |
+ *                 +------+       +-----+
+ */
+
+#include <stdint.h>
+#include "chan.h"
+#include "mux.h"
+#include "sort.h"
+
+struct breakdown_cpu {
+	struct mux  mux;
+	struct chan tri;
+};
+
+struct breakdown_emu {
+	int64_t nphycpus;
+	struct sort sort;
+	struct pvt *pvt;
+};
+
+#endif /* BREAKDOWN_H */

src/emu/openmp/openmp_priv.h

@@ -8,6 +8,7 @@
 #include "task.h"
 #include "model_cpu.h"
 #include "model_thread.h"
+#include "breakdown.h"
 
 /* Private enums */
 
@@ -63,6 +64,7 @@ struct openmp_thread {
 
 struct openmp_cpu {
 	struct model_cpu m;
+	struct breakdown_cpu breakdown;
 };
 
 struct openmp_proc {
@@ -70,10 +72,19 @@ struct openmp_proc {
 	struct task_info task_info;
 };
 
+struct openmp_emu {
+	struct breakdown_emu breakdown;
+};
+
 int model_openmp_probe(struct emu *emu);
 int model_openmp_create(struct emu *emu);
 int model_openmp_connect(struct emu *emu);
 int model_openmp_event(struct emu *emu);
 int model_openmp_finish(struct emu *emu);
 
+int model_openmp_breakdown_create(struct emu *emu);
+int model_openmp_breakdown_connect(struct emu *emu);
+int model_openmp_breakdown_finish(struct emu *emu,
+		const struct pcf_value_label **labels);
+
 #endif /* OPENMP_PRIV_H */

src/emu/openmp/setup.c

@@ -278,6 +278,19 @@ model_openmp_create(struct emu *emu)
 		}
 	}
 
+	struct openmp_emu *e = calloc(1, sizeof(struct openmp_emu));
+	if (e == NULL) {
+		err("calloc failed:");
+		return -1;
+	}
+
+	extend_set(&emu->ext, model_id, e);
+
+	if (model_openmp_breakdown_create(emu) != 0) {
+		err("model_openmp_breakdown_create failed");
+		return -1;
+	}
+
 	return 0;
 }
 
@@ -294,6 +307,11 @@ model_openmp_connect(struct emu *emu)
 		return -1;
 	}
 
+	if (model_openmp_breakdown_connect(emu) != 0) {
+		err("model_openmp_breakdown_connect failed");
+		return -1;
+	}
+
 	return 0;
 }
 
@@ -379,6 +397,11 @@ model_openmp_finish(struct emu *emu)
 		return -1;
 	}
 
+	if (model_openmp_breakdown_finish(emu, pcf_labels) != 0) {
+		err("model_openmp_breakdown_finish failed");
+		return -1;
+	}
+
 	/* When running in linter mode perform additional checks */
 	if (emu->args.linter_mode && end_lint(emu) != 0) {
 		err("end_lint failed");

test/rt/openmp/CMakeLists.txt

@@ -26,7 +26,7 @@ if(NOT NOSV_FOUND)
 endif()
 
 function(openmp_rt_test)
-  ovni_test(${ARGN})
+  ovni_test(${ARGN} SORT)
   target_compile_options("${OVNI_TEST_NAME}" PUBLIC "-fopenmp=libompv"
 	  "-no-pedantic")
   target_link_options("${OVNI_TEST_NAME}" PUBLIC "-fopenmp=libompv")
@@ -34,7 +34,7 @@ function(openmp_rt_test)
   set_property(TEST "${OVNI_TEST_NAME}" APPEND PROPERTY
     ENVIRONMENT "OMP_OVNI=1")
   set_property(TEST "${OVNI_TEST_NAME}" APPEND PROPERTY
-    ENVIRONMENT "NOSV_CONFIG_OVERRIDE=instrumentation.version=ovni")
+    ENVIRONMENT "NOSV_CONFIG_OVERRIDE=instrumentation.version=ovni,ovni.level=3")
 endfunction()
 
 openmp_rt_test(barrier-explicit.c)
@@ -53,7 +53,7 @@ openmp_rt_test(taskloop.c)
 openmp_rt_test(taskwait.c)
 openmp_rt_test(team-distribute.c)
 openmp_rt_test(worksharing-and-tasks.c)
-openmp_rt_test(worksharing-mix.c)
+openmp_rt_test(worksharing-mix.c BREAKDOWN)
 openmp_rt_test(worksharing-task.c)
 openmp_rt_test(worksharing.c)
 openmp_rt_test(worksharing01.c)

test/rt/openmp/worksharing-mix.c

@@ -34,7 +34,7 @@ int main(void)
 		#pragma omp single
 		for (int i = 0; i < 100; i++)
 		{
-			#pragma omp task
+			#pragma omp task label("mini-task")
 			sleep_us(10);
 		}
 	}